Experimenting with machine learning in R with tidymodels and the Kaggle titanic dataset

1 Motivation

I would like to familiarize myself with machine learning (ML) techniques in R. So I have been reading and learning by doing. I thought I’d share my experience for others who’d like to give it a try.

The two great books I’m using are:

• An Introduction to Statistical Learning with Applications in R by Gareth James, Daniela Witten, Trevor Hastie and Robert Tibshirani

• Tidy models in R by Max Kuhn and Julia Silge

I also recommend checking out the material (codes, screencasts) shared by David Robinson and Julia Silge from whom I picked some useful tricks that I put to use below.

To try things, I’ve joined the Kaggle online community which gathers folks with lots of experience in ML from whom you can learn. Kaggle also hosts public datasets that can be used for playing around.

Let’s start with the famous Titanic dataset. We need to predict if a passenger survived the sinking of the Titanic (1) or not (0). A dataset is provided for training our models (train.csv). Another dataset is provided (test.csv) for which we do not know the answer. We will predict survival for each passenger, submit our answer to Kaggle and see how well we did compared to other folks. The metric for comparison is the percentage of passengers we correctly predict – aka as accuracy.

First things first, let’s load some packages to get us started.

library(tidymodels) # metapackage for ML 
library(tidyverse) # metapackage for data manipulation and visulaisation
library(stacks) # stack ML models for better perfomance
theme_set(theme_light())
doParallel::registerDoParallel(cores = 4) # parallel computations

2 Data

Read in training data.

rawdata <- read_csv("dat/titanic/train.csv")
glimpse(rawdata)

## Rows: 891
## Columns: 12
## $ PassengerId <dbl> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,…
## $ Survived    <dbl> 0, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 1, 0, 1, 0, 1…
## $ Pclass      <dbl> 3, 1, 3, 1, 3, 3, 1, 3, 3, 2, 3, 1, 3, 3, 3, 2, 3, 2, 3, 3…
## $ Name        <chr> "Braund, Mr. Owen Harris", "Cumings, Mrs. John Bradley (Fl…
## $ Sex         <chr> "male", "female", "female", "female", "male", "male", "mal…
## $ Age         <dbl> 22, 38, 26, 35, 35, NA, 54, 2, 27, 14, 4, 58, 20, 39, 14, …
## $ SibSp       <dbl> 1, 1, 0, 1, 0, 0, 0, 3, 0, 1, 1, 0, 0, 1, 0, 0, 4, 0, 1, 0…
## $ Parch       <dbl> 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 1, 0, 0, 5, 0, 0, 1, 0, 0, 0…
## $ Ticket      <chr> "A/5 21171", "PC 17599", "STON/O2. 3101282", "113803", "37…
## $ Fare        <dbl> 7.2500, 71.2833, 7.9250, 53.1000, 8.0500, 8.4583, 51.8625,…
## $ Cabin       <chr> NA, "C85", NA, "C123", NA, NA, "E46", NA, NA, NA, "G6", "C…
## $ Embarked    <chr> "S", "C", "S", "S", "S", "Q", "S", "S", "S", "C", "S", "S"…

naniar::miss_var_summary(rawdata)

After some data exploration (not shown), I decided to take care of missing values, gather the two family variables in a single variable, and create a variable title.

# Get most frequent port of embarkation
uniqx <- unique(na.omit(rawdata$Embarked))
mode_embarked <- as.character(fct_drop(uniqx[which.max(tabulate(match(rawdata$Embarked, uniqx)))]))

# Build function for data cleaning and handling NAs
process_data <- function(tbl){
  
  tbl %>%
    mutate(class = case_when(Pclass == 1 ~ "first",
                             Pclass == 2 ~ "second",
                             Pclass == 3 ~ "third"),
           class = as_factor(class),
           gender = factor(Sex),
           fare = Fare,
           age = Age,
           alone = if_else(SibSp + Parch == 0, "yes", "no"), # alone variable
           alone = as_factor(alone),
           port = factor(Embarked), # rename embarked as port
           title = str_extract(Name, "[A-Za-z]+\\."), # title variable
           title = fct_lump(title, 4)) %>% # keep only most frequent levels of title
    mutate(port = ifelse(is.na(port), mode_embarked, port), # deal w/ NAs in port (replace by mode)
           port = as_factor(port)) %>%
    group_by(title) %>%
    mutate(median_age_title = median(age, na.rm = T)) %>%
    ungroup() %>%
    mutate(age = if_else(is.na(age), median_age_title, age)) # deal w/ NAs in age (replace by median in title)
}

# Process the data
dataset <- rawdata %>%
  process_data() %>%
  mutate(survived = as_factor(if_else(Survived == 1, "yes", "no"))) %>%
  select(survived, class, gender, age, alone, fare, port, title) 

# Have a look again
glimpse(dataset)

## Rows: 891
## Columns: 8
## $ survived <fct> no, yes, yes, yes, no, no, no, no, yes, yes, yes, yes, no, no…
## $ class    <fct> third, first, third, first, third, third, first, third, third…
## $ gender   <fct> male, female, female, female, male, male, male, male, female,…
## $ age      <dbl> 22, 38, 26, 35, 35, 30, 54, 2, 27, 14, 4, 58, 20, 39, 14, 55,…
## $ alone    <fct> no, no, yes, no, yes, yes, yes, no, no, no, no, yes, yes, no,…
## $ fare     <dbl> 7.2500, 71.2833, 7.9250, 53.1000, 8.0500, 8.4583, 51.8625, 21…
## $ port     <fct> 3, 1, 3, 3, 3, 2, 3, 3, 3, 1, 3, 3, 3, 3, 3, 3, 2, 3, 3, 1, 3…
## $ title    <fct> Mr., Mrs., Miss., Mrs., Mr., Mr., Mr., Master., Mrs., Mrs., M…

naniar::miss_var_summary(dataset)

Let’s apply the same treatment to the test dataset.

rawdata <- read_csv("dat/titanic/test.csv") 
holdout <- rawdata %>%
  process_data() %>%
  select(PassengerId, class, gender, age, alone, fare, port, title) 

glimpse(holdout)

## Rows: 418
## Columns: 8
## $ PassengerId <dbl> 892, 893, 894, 895, 896, 897, 898, 899, 900, 901, 902, 903…
## $ class       <fct> third, third, second, third, third, third, third, second, …
## $ gender      <fct> male, female, male, male, female, male, female, male, fema…
## $ age         <dbl> 34.5, 47.0, 62.0, 27.0, 22.0, 14.0, 30.0, 26.0, 18.0, 21.0…
## $ alone       <fct> yes, no, yes, yes, no, yes, yes, no, yes, no, yes, yes, no…
## $ fare        <dbl> 7.8292, 7.0000, 9.6875, 8.6625, 12.2875, 9.2250, 7.6292, 2…
## $ port        <fct> 2, 3, 2, 3, 3, 3, 2, 3, 1, 3, 3, 3, 3, 3, 3, 1, 2, 1, 3, 1…
## $ title       <fct> Mr., Mrs., Mr., Mr., Mrs., Mr., Miss., Mr., Mrs., Mr., Mr.…

naniar::miss_var_summary(holdout)

3 Exploratory data analysis

skimr::skim(dataset)

Data summary

Name	dataset
Number of rows	891
Number of columns	8
_______________________
Column type frequency:
factor	6
numeric	2
________________________
Group variables	None

Variable type: factor

skim_variable	n_missing	complete_rate	ordered	n_unique	top_counts
survived	0	1	FALSE	2	no: 549, yes: 342
class	0	1	FALSE	3	thi: 491, fir: 216, sec: 184
gender	0	1	FALSE	2	mal: 577, fem: 314
alone	0	1	FALSE	2	yes: 537, no: 354
port	0	1	FALSE	4	3: 644, 1: 168, 2: 77, S: 2
title	0	1	FALSE	5	Mr.: 517, Mis: 182, Mrs: 125, Mas: 40

Variable type: numeric

skim_variable	n_missing	complete_rate	mean	sd	p0	p25	p50	p75	p100	hist
age	0	1	29.39	13.26	0.42	21.00	30.00	35	80.00	▂▇▃▁▁
fare	0	1	32.20	49.69	0.00	7.91	14.45	31	512.33	▇▁▁▁▁

dataset %>%
  group_by(gender) %>%
  summarize(n = n(),
            n_surv = sum(survived == "yes"),
            pct_surv = n_surv / n)

dataset %>%
  group_by(title) %>%
  summarize(n = n(),
            n_surv = sum(survived == "yes"),
            pct_surv = n_surv / n) %>%
  arrange(desc(pct_surv))

dataset %>%
  group_by(class, gender) %>%
  summarize(n = n(),
            n_surv = sum(survived == "yes"),
            pct_surv = n_surv / n) %>%
  arrange(desc(pct_surv))

dataset %>%
  group_by(class, gender) %>%
  summarize(n = n(),
            n_surv = sum(survived == "yes"),
            pct_surv = n_surv / n) %>%
    mutate(class = fct_reorder(class, pct_surv)) %>%
    ggplot(aes(pct_surv, class, fill = class, color = class)) +
    geom_col(position = position_dodge()) +
    scale_x_continuous(labels = percent) +
    labs(x = "% in category that survived", fill = NULL, color = NULL, y = NULL) +
  facet_wrap(~gender)

dataset %>%
  mutate(age = cut(age, breaks = c(0, 20, 40, 60, 80))) %>%
  group_by(age, gender) %>%
  summarize(n = n(),
            n_surv = sum(survived == "yes"),
            pct_surv = n_surv / n) %>%
    mutate(age = fct_reorder(age, pct_surv)) %>%
    ggplot(aes(pct_surv, age, fill = age, color = age)) +
    geom_col(position = position_dodge()) +
    scale_x_continuous(labels = percent) +
    labs(x = "% in category that survived", fill = NULL, color = NULL, y = NULL) +
  facet_wrap(~gender)

dataset %>%
    ggplot(aes(fare, group = survived, color = survived, fill = survived)) +
    geom_histogram(alpha = .4, position = position_dodge()) +
    labs(x = "fare", y = NULL, color = "survived?", fill = "survived?")

4 Gradient boosting algorithms

Let’s start with gradient boosting methods which are very population in the ML community.

4.1 Training/testing datasets

Split our dataset in two, one dataset for training and the other one for testing. We will use an additionnal splitting step for cross-validation.

set.seed(2021)
spl <- initial_split(dataset)
train <- training(spl)
test <- testing(spl)

train_5fold <- train %>%
  vfold_cv(5)

4.2 Tuning

Set up defaults.

mset <- metric_set(accuracy) # metric is accuracy
control <- control_grid(save_workflow = TRUE,
                        save_pred = TRUE,
                        extract = extract_model) # grid for tuning

First a recipe.

xg_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)

Then specify a gradient boosting model.

xg_model <- boost_tree(mode = "classification", # binary response
                       trees = tune(),
                       mtry = tune(),
                       tree_depth = tune(),
                       learn_rate = tune()) # parameters to be tuned

Now set our workflow.

xg_wf <- 
  workflow() %>% 
  add_model(xg_model) %>% 
  add_recipe(xg_rec)

Use cross-validation to evaluate our model with different param config.

xg_tune <- xg_wf %>%
  tune_grid(train_5fold,
            metrics = mset,
            control = control,
            grid = crossing(trees = seq(500, 2500, 500),
                            mtry = c(3, 5, 8), # finalize(mtry(), train)
                            tree_depth = c(5, 10, 15),
                            learn_rate = c(0.01, 0.005)))

Visualize the results.

autoplot(xg_tune)

Collect metrics.

xg_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

4.3 Fit model

Use best config to fit model to training data.

xg_fit <- xg_wf %>%
  finalize_workflow(select_best(xg_tune)) %>%
  fit(train)

## [14:52:06] WARNING: amalgamation/../src/learner.cc:1095: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

Check out accuracy on testing dataset to see if we overfitted.

xg_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

Check out important features (aka predictors).

importances <- xgboost::xgb.importance(model = extract_fit_engine(xg_fit))
importances %>%
  mutate(Feature = fct_reorder(Feature, Gain)) %>%
  ggplot(aes(Gain, Feature)) +
  geom_col()

4.4 Make predictions

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

xg_fit %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/xgboost.csv")

I got and accuracy of 0.76794. Cool. Let’s train a random forest model now.

5 Random forests

Let’s continue with random forest methods.

5.1 Tuning

First a recipe.

rf_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_dummy(all_nominal_predictors()) # all factors var are split into binary terms (factor disj coding)

Then specify a gradient boosting model.

rf_model <- rand_forest(mode = "classification", # binary response
                        engine = "ranger", # by default
                        mtry = tune(),
                        trees = tune(),
                        min_n = tune()) # parameters to be tuned

Now set our workflow.

rf_wf <- 
  workflow() %>% 
  add_model(rf_model) %>% 
  add_recipe(rf_rec)

Use cross-validation to evaluate our model with different param config.

rf_tune <- rf_wf %>%
  tune_grid(train_5fold,
            metrics = mset,
            control = control,
            grid = crossing(trees = seq(500, 2500, 500),
                            mtry = c(3, 5, 8), # finalize(mtry(), train)
                            min_n = seq(1, 20, 5)))

Visualize the results.

autoplot(rf_tune)

Collect metrics.

rf_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

5.2 Fit model

Use best config to fit model to training data.

rf_fit <- rf_wf %>%
  finalize_workflow(select_best(rf_tune)) %>%
  fit(train)

Check out accuracy on testing dataset to see if we overfitted.

rf_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

Check out important features (aka predictors).

library(vip)
finalize_model(
  x = rf_model,
  parameters = select_best(rf_tune)) %>%
  set_engine("ranger", importance = "permutation") %>%
  fit(survived ~ ., data = juice(prep(rf_rec))) %>%
  vip(geom = "point")

5.3 Make predictions

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

rf_fit %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/randomforest.csv")

I got and accuracy of 0.77033, a bit better than gradient boosting.

6 Regularization methods (elastic net here)

Let’s continue with elastic net regularization.

6.1 Tuning

First a recipe.

en_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_dummy(all_nominal_predictors()) %>% # all factors var are split into binary terms (factor disj coding)
  step_normalize(all_predictors()) # normalize

Then specify a gradient boosting model.

en_model <- logistic_reg(penalty = tune()) %>% # penalty to be tuned
  set_engine("glmnet") %>% # elastic net
  set_mode("classification") # binary response

Now set our workflow.

en_wf <- 
  workflow() %>% 
  add_model(en_model) %>% 
  add_recipe(en_rec)

Use cross-validation to evaluate our model with different param config.

en_tune <- en_wf %>%
  tune_grid(train_5fold,
            metrics = mset,
            control = control,
            grid = crossing(penalty = 10 ^ seq(-8, -.5, .5)))

Visualize the results.

autoplot(en_tune)

Collect metrics.

en_tune %>%
  collect_metrics() %>%
  arrange(desc(mean))

6.2 Fit model

Use best config to fit model to training data.

en_fit <- en_wf %>%
  finalize_workflow(select_best(en_tune)) %>%
  fit(train)

Check out accuracy on testing dataset to see if we overfitted.

en_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

Check out important features (aka predictors).

library(broom)
en_fit$fit$fit$fit %>%
  tidy() %>%
  filter(lambda >= select_best(en_tune)$penalty) %>%
  filter(lambda == min(lambda),
         term != "(Intercept)") %>%
  mutate(term = fct_reorder(term, estimate)) %>%
  ggplot(aes(estimate, term, fill = estimate > 0)) +
  geom_col() +
  theme(legend.position = "none")

6.3 Make predictions

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

en_fit %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/elasticnet.csv")

I got and accuracy of 0.76794.

7 Stacked ensemble modelling

Tune again with a probability-based metric.

xg_tune <- xg_wf %>%
  tune_grid(train_5fold,
            metrics = metric_set(roc_auc),
            control = control,
            grid = crossing(trees = seq(500, 2500, 500),
                            mtry = c(3, 5, 8), # finalize(mtry(), train)
                            tree_depth = c(5, 10, 15),
                            learn_rate = c(0.01, 0.005)))

rf_tune <- rf_wf %>%
  tune_grid(train_5fold,
            metrics = metric_set(roc_auc),
            control = control,
            grid = crossing(trees = seq(500, 2500, 500),
                            mtry = c(3, 5, 8), # finalize(mtry(), train)
                            min_n = seq(1, 20, 5)))

en_tune <- en_wf %>%
  tune_grid(train_5fold,
            metrics = metric_set(roc_auc),
            control = control,
            grid = crossing(penalty = 10 ^ seq(-8, -.5, .5)))

Get best config.

xg_best <- xg_tune %>% filter_parameters(parameters = select_best(xg_tune))
rf_best <- rf_tune %>% filter_parameters(parameters = select_best(rf_tune))
en_best <- en_tune %>% filter_parameters(parameters = select_best(en_tune))

Do the stacked ensemble modelling (for some reasons I didn’t understand, I could not stack elastic net).

blended <- stacks() %>% # initialize
  add_candidates(xg_best) %>% # add gradient boosting
  add_candidates(rf_best) %>% # add random forest
#  add_candidates(en_best) %>% # add elastic net
  blend_predictions() %>% # fit regularized model
  fit_members() # fit candidate members with non-zero stacking coef with full training dataset

Perf on testing dataset?

test %>%
  bind_cols(predict(blended, .)) %>%
  accuracy(survived, .pred_class)

Now predict.

holdout %>%
  bind_cols(predict(blended, .)) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/stacked.csv")

I got an 0.77033 accuracy.

8 Logistic regression

And what about a good old-fashioned logistic regression (not a ML algo)?

First a recipe.

logistic_rec <- recipe(survived ~ ., data = train) %>%
  step_impute_median(all_numeric()) %>% # replace missing value by median
  step_dummy(all_nominal_predictors()) %>% # all factors var are split into binary terms (factor disj coding)
  step_normalize(all_predictors()) # normalize

Then specify a logistic regression.

logistic_model <- logistic_reg() %>% # no param to be tuned
  set_engine("glm") %>% # elastic net
  set_mode("classification") # binary response

Now set our workflow.

logistic_wf <- 
  workflow() %>% 
  add_model(logistic_model) %>% 
  add_recipe(logistic_rec)

Fit model.

logistic_fit <- logistic_wf %>%
  fit(train)

Inspect results.

tidy(logistic_fit, exponentiate = TRUE) %>%
  filter(p.value < 0.05)

Check out accuracy on testing dataset to see if we overfitted.

logistic_fit %>%
  augment(test, type.predict = "response") %>%
  accuracy(survived, .pred_class)

Confusion matrix.

logistic_fit %>%
  augment(test, type.predict = "response") %>%
  conf_mat(survived, .pred_class)

##           Truth
## Prediction  no yes
##        no  133  22
##        yes  17  51

Custom metrics.

custom_metrics <- metric_set(sens, precision, recall, f_meas)
logistic_fit %>%
  augment(test, type.predict = "response") %>%
  custom_metrics(truth = survived, estimate = .pred_class)

ROC curve.

logistic_fit %>%
  augment(test, type.predict = "response") %>%
  roc_curve(truth = survived, estimate = .pred_yes, event_level = "second") %>%
  autoplot()

Check out important features (aka predictors).

library(broom)
logistic_fit %>%
  tidy() %>%
  mutate(term = fct_reorder(term, estimate)) %>%
  ggplot(aes(estimate, term, fill = estimate > 0)) +
  geom_col() +
  theme(legend.position = "none")

Now we’re ready to predict survival for the holdout dataset and submit to Kaggle.

logistic_fit %>%
  augment(holdout) %>%
  select(PassengerId, Survived = .pred_class) %>%
  mutate(Survived = if_else(Survived == "yes", 1, 0)) %>%
  write_csv("output/titanic/logistic.csv")

I got and accuracy of 0.76555. Oldies but goodies!

9 Conclusions

I covered three ML algorithms (gradient boosting, random forest and elastic net) and logistic regression with the awesome tidymodels metapackage in R. My scores at predicting Titanic survivors were ok I guess. Some folks on Kaggle got a perfect accuracy, so there is always room for improvement. Maybe better tuning, better features (or predictors) or other algorithms would increase accuracy. Of course, I forgot to use set.seed() so results are not exactly reproducible.